PIQL: Success- Tolerant Query Processing in the Cloud

Michael Armbrust, Kristal Curtis, Tim KraskaArmando Fox, Michael J. Franklin, David A.

PattersonAMP Lab, EECS, UC Berkeley

Introduction

• Large-scale websites are increasingly moving from relational databases to distributed key-value stores.

• Why? - High request rate- Low latency workloads- Scalability

Key-value stores at a cost of ?

• Writing complex imperative functions• Index management• Intra query parallelization• And LOSS of DATA INDEPENCE

A blend of both: PIQL

• Performance predictable subset of SQL• Benefits of RDBMS such as ability to express

queries declaratively• Physical data independence• Automatic index selection and maintenance• Real time guarantees on PERFORMANCE that

come from underlying key-value store

Features:

• Run on the top of key/value stores• Bounds on the number of operations that will

be performed on key-value store• Compile time feedback on worst-case

performance for all queries• Automatic selection and maintenance of

indexes

Alternative approach

• Complex developer written imperative programs• Example:

Data model of Cassandra For a query: search for messages that contain

certain wordValues inserted of the form:row -> userid, supercolumn ->word, column ->messageTimestamp, value->messageId

Equivalent PIQL query:

FETCH messageOF user BY recipientWHERE user = [this] ANDmessage.text CONTAINS [1: word]ORDER BY timestamp

Query Scaling classes

• Class 1: (Constant)- Amount of data required to process the query is constant

• Class 2: (Bounded)- Amount of data required to process the query is naturally bounded

• Class 3: (Sub-linear or Linear)- Amount of data required to process the query grows sub-linearly eventually

• Class 4: (Super-linear)-

PIQL Query Syntax

• name• Parameters [ordinal:name]General syntax:QUERY nameFETCH entity[OF joined-entity alias BY relationship] ...

WHERE predicates[{PAGINATE perpage | LIMIT count}]

Example Class 1

• To return profile of a user given a user nameQUERY userByNameFETCH userWHERE user.name = [1:name]

• Calculating bound:Simple: 1 or zero results

Example Class 2

• To return users by their hometownQUERY userByHometownFETCH userWHERE user.hometown = [1:hometown]LIMIT [1:count] MAX 100

• Calculating bound:• LIMIT clause returns at most 100 items

Example Class 3

• To return a list of the most recent thoughts owned by a particular user

QUERY userThoughtsFETCH thought ofuser by ownerWHERE user.name = [1:username]ORDER BY timestampLIMIT [2:count] MAX 100Bound: 100

Example Class 4

To return a paginated list, 10 at a time, of the most recent thoughts of all the approved subscriptions owned by the current user.

QUERY thoughtstreamFETCH thoughtOF user friend BY ownerOF subscription BY targetOF user me BY ownerWHERE me.username=[1:username] AND approved = true

ORDER BY timestampPAGINATE 10

Comparison

• Graph

Queries in PIQL

• Entities analogous to Relations• Queries are specified as templates ahead of

time• No of operations required in worst case are

provided to developer as feedback at compile time

Architecture Overview

Optimization in PIQL

• Phase 1-(Stop Operator Insertion)

Optimization in PIQL

• Phase 2

Prediction Framework

Performance Insight Assistant• Provides feedback to developer to fix ‘unsafe’

queries• Guidance on how to set a ‘Cardinality limit’

compatible with SLO Compliance• Provides a chart of latency distribution for

each setting of the cardinality

Performance Insight assistant

• Predicted Heat Map for Thoughtstream query

Execution Engine

• Leverages key-value store to achieve scalability and high performance

• Requests to a key-value store are done in parallel

• Limit hint information is used to prefetch all required data in single request

Performance overview

Fig: System scaling in number of users/machines with constant query latency

Conclusion

• Performance predictability and scalability of Key-value stores + Scale independence of Relational Model= PIQL

• GQL, HIVE, PIG, VoltDB are also on similar grounds but they are focused on Batch Analytics rather than Interactive applications